Not Your Weights,
Not Your Product? A Sequoia Partner's Warning

The biggest myth about Sovereign AI: it's not about ditching APIs

In early 2026, Sequoia Capital partner Sonya Huang made a striking claim at a public event: the competitive battlefield in AI has already sunk from the application layer down to the "intelligence layer" — if enterprises don't own their model weights, they don't really own their product. A year earlier, most people would have found this claim baffling, since the prevailing consensus was still that "large models are a game for tech giants; startups should just build a good wrapper on top."

But she immediately made an important clarification: this isn't asking enterprises to abandon closed APIs like OpenAI or Anthropic, nor does it require building a model from scratch through pre-training. What she calls "Sovereign AI" is really about vertically integrating the most critical functions — keeping the layer of intelligence that actually drives product differentiation in your own hands, rather than outsourcing all of your intelligence capability. This isn't a binary choice; it's about having the ability to clearly define which parts can be outsourced and which parts must stay under your own control.

This also touches on a more fundamental fork in the road: will the future of AI be driven by a handful of centralized, all-powerful models, or will it be built on open, capable foundation models on top of which every enterprise and vertical builds its own customized intelligence? Sonya Huang leans toward the latter — because that's the healthier ecosystem, one where every company can grow its own differentiated moat, instead of every application looking like it came out of the same mold.

The four drivers behind enterprises building their own models

Why is 2026 the moment when more and more application companies are seriously considering building their own models and owning their weights? Sonya Huang identifies four core drivers:

1. Cost — API pricing is a linear, usage-based cost: the more users and calls you have, the bigger the bill, and the vendor can raise prices whenever it wants, with the enterprise having zero control. Once you own your model, the marginal cost of inference after training is low and gets diluted further with scale — for high-frequency calls or large-scale deployment, the long-term cost of serving is far lower than continuously calling an API.
2. Speed — This isn't just about response latency. A small, specialized model trained for a specific domain — like code completion or log-anomaly detection — doesn't need to cover general-purpose capabilities, so all of its parameter capacity goes toward a single task. That can make inference several times faster than a general-purpose large model, which makes a decisive difference to user experience in real-time interaction scenarios.
3. Performance — This is the most counterintuitive point: as of 2026, open-weight models are rapidly closing the gap with closed frontier models, and when combined with post-training and fine-tuning on an enterprise's proprietary data, they can even outperform general-purpose closed APIs in vertical domains. The reason is straightforward: an enterprise understands its own business context better than any large lab ever could, and that proprietary data is something a general-purpose model will never see.
4. Ownership of your own destiny — If your core intelligence depends entirely on a single external vendor, the real leverage over your business sits with someone else. If the vendor raises prices, changes its interface, throttles usage, or even shuts down the service, the enterprise has no recourse — and for regulated sectors like finance, healthcare, and government, there's also the compliance and security risk of user data flowing upstream to a third party.

It's worth noting that these four drivers don't mean every company needs to build its own model — they're the criteria for drawing the line between "what capability must stay in-house" and "what can be outsourced."

Not Your Weights, Not Your Product: weights are AI's private key

Sonya Huang delivered the single most important line of the entire talk: "Not your weights, not your product." The phrase is adapted from the classic crypto maxim "not your keys, not your coins" — coins sitting on an exchange aren't really yours, because the exchange holds the private key, and it can freeze your assets or run into any number of problems. Only when you hold the private key yourself do the coins truly belong to you.

The sting in that line is this: if your product is just a wrapper calling an API, every UX improvement you ship today can be wiped out tomorrow by a single upstream model update. And all the data and domain know-how you accumulate along the way never truly becomes your moat unless it's distilled into a model you actually own.

Application companies are the new AI labs: a four-step playbook

At this point, many application companies' first reaction is: "Building models is what AI labs do — we're an applications company, how could we possibly pull that off?" Sonya Huang's view is exactly the opposite: today's application companies are, in essence, the new generation of AI labs. Look at companies like Harvey in legal AI, Factory in intelligent manufacturing, or Glean in enterprise search — their teams all have strong applied-research capability. They're doing fine-tuning, post-training, inference optimization, and building their own evaluation systems — they're not simply wrapping an API.

For a company with zero prior model experience, she lays out a practical four-step path:

1. Define the build-versus-rent boundary first — This is a business decision, not a technical one. Weigh cost, speed and latency, performance, and the value of your proprietary data together, and draw a clear line: build what's on this side of the line yourself, and confidently rent everything on the other side.
2. Stand up a dedicated, agile frontier-research team — Don't let your existing platform team squeeze this in as a side project — that produces something half-baked. Stand up a small, dedicated team of a few people whose KPI is improving model performance and maximizing the value of your data, not shipping engineering artifacts.
3. Build external visibility (a technical brand) — Publish research findings, technical blog posts, and open-source contributions to build your reputation. This isn't just about attracting top AI talent — it's also what convinces customers and partners that you genuinely have the technical capability you claim to have.
4. Design a clear technical roadmap — Establish explicit model-evaluation standards, decide on your router and application framework, and leave architectural room to progressively move upstream — from prompt engineering to post-training, and eventually pre-training if needed. Lay a solid enough foundation, but you don't have to build the skyscraper on day one.

The modern enterprise AI stack: production and development in parallel

Sonya Huang also sketched out what a typical modern enterprise AI stack looks like today — really two environments running in parallel. The production environment faces users and is optimized for stability, low latency, and high availability. The development environment covers model evaluation, data processing, and the training and evolution engine. The two must be connected: better models iterated in the development environment need to deploy safely and smoothly into production — otherwise the model stalls the moment it ships, and performance never moves past day one.

She specifically pointed out what makes this era different from just two years ago: back then, if you wanted GPT-4-level performance, calling OpenAI's API was pretty much your only option. Today, open-weight foundation models like Kimi K3 and Zhipu's GLM-5.2, combined with post-training, prompt engineering, and RAG, can already match or exceed the performance of closed APIs — and their weights are fully open, so an enterprise can take them in-house and modify them however it wants. The core insight of the whole talk is this: the competitive moat in AI keeps sinking lower — from the chat interface, to prompting, to workflow orchestration, all of it is being commoditized fast, and the real moat is settling into intelligence itself, which is to say, the model weights.

This article is compiled from a public talk by Sequoia Capital partner Sonya Huang, as summarized by the YouTube channel "最佳拍檔" (hosted by Dafei). It does not constitute a guarantee or commitment of any kind — please evaluate your organization's actual adoption strategy based on your own circumstances.

SecuAgent's take: own control first, own weights later

Sonya Huang's argument is precise in a Silicon Valley context — but when applied to most SMEs and traditional industries in Taiwan, SecuAgent thinks one key gap needs to be named up front: what most enterprises actually need isn't literally "owning model weights" — it's "owning control over the intelligence layer." The two are highly correlated, but they are not the same thing.

Building your own training pipeline, hiring a frontier-research team, and standing up an evaluation system is still a high bar for a traditional manufacturer or service business with a few hundred employees and a limited AI budget — and that's exactly where the four dimensions from the talk (cost, speed, performance, proprietary data) should be used to draw the line. For most Taiwanese enterprises, the first step doesn't need to be pre-training or large-scale post-training — it's making three practical pieces of "control" solid first:

• Data governance before model-building — If your proprietary data hasn't first been desensitized, classified, and structured, feeding it into a model you build or fine-tune later just feeds it noise instead. The first layer of control is owning your own data pipeline — not rushing to swap models.
• Use routing and guardrails to buy flexibility without lock-in — An enterprise doesn't need to own weights from day one to dynamically route tasks across different models — open-weight, closed, or fine-tuned in-house — while a dual-model safety guardrail keeps sensitive data contained at the edge. Earn the freedom to switch vendors first, then decide whether to move further upstream toward building your own.
• Replace "it feels right" with an evaluation system — The step in Sonya Huang's four-step path most likely to get skipped by Taiwanese enterprises is "establish explicit model-evaluation standards." Without an evaluation system, swapping models, fine-tuning, and post-training are all guesswork — you can't prove the investment paid off, let alone justify moving further upstream.

Put another way: the Taiwan-market version of "not your weights, not your product" might read closer to "not your control, not your AI strategy." You don't necessarily need to build your own power plant — but you do need your own water meter and electricity meter, so you know how intelligence is being used, by whom, and where it's flowing. That's the "weight-level" agency that most enterprises can actually afford — and actually use — right now.